TRA-1-4 PS2-54 OPTIMIZING CONJOINT ANALYSIS IDENTIFICATION OF SUBGROUP PREFERENCES: A NOVEL APPROACH TO ENHANCING SHARED DECISION MAKING

Monday, October 24, 2016: 10:45 AM
Bayshore Ballroom Salon D, Lobby Level (Westin Bayshore Vancouver)
Poster Board # PS2-54

Azza Shoaibi1, Brian Neelon2 and Leslie Lenert, MD2, (1)Medical University of South Carolina, charleston, SC, (2)CHARLESTON, SC
Purpose: Conjoint analysis (CA) is a preference elicitation method used to characterize the utility that individuals ascribe to dimensions of care. An aspect of CA of increasing importance is market segmentation—that is, finding different subgroups in the population in order to tailor options to group-level preferences. Segmentation analyses typically adopt a two-step process that may lead to biased inferences. We present a novel approach optimized specifically to accurately identify relevant population segments for shared decision making.

Method: We extended the classic CA model by embedding the model within a Bayesian nonparametric mixture framework that can accurately predict individual-level utilities and optimally cluster patients in a single step. A stick-breaking Dirichlet process (DP) was used to define an infinite CA mixture model. We extended the DP model to incorporate covariate information not only in the cluster means but also in the cluster probabilities by adopting a probit stick-breaking process (PSBP).

We conducted extensive simulation studies to evaluate model performance, assess Markov chain Monte Carlo (MCMC) convergence, and compare our approach to existing methods for CA. We implemented the simulations by generating multiple datasets under the assumed model (e.g., a CA mixture model with a fixed number of clusters), and on each dataset, we computed estimates of parameters and other quantities of interest. We quantified the model’s performance in terms of parameter bias, mean square error (MSE), and coverage probabilities, all summarized as Monte Carlo averages over the simulated datasets. We used the RAND and Jaccard similarity indices to compare the true and model-predicted cluster allocations. We also compared our method to existing two-stage approaches (a hierarchical Bayes model for generating individual partworth predictions followed by a latent class analysis).

Results:   To date, we have derived mathematical results that provide the theoretical foundation for our approach. Simulation studies are currently underway to test and validate the method. These simulations indicate successful and efficient convergence of the model, with similarity indices indicating accurate clustering. In addition, initial results show that the developed model improved true classification by approximately 20% when compared to conventional two-stage methods.  

Conclusions:  Integrating CA and segmentation is a promising approach for measuring patient preferences and tailoring treatment recommendations to patient subpopulatio